BiSPD-YOLO: Surface Defect Detection Method for Small Features and Low-resolution Images

At present, deep learning objective detection method based on learning features suffer from low detection rates and poor accuracy rates in metal surface defect detection. This is primarily due to the fact that the detected images are mostly gray images with small features and low resolution, which makes the model inefficient to train and slow to converge. This paper proposes a BiSPD-YOLO metal surface defect detection model based on YOLOv5 to solve these problems. Firstly, this model uses SDP-Conv module to replace the traditional strided convolution and pooling to enhance the training of the network for low-resolution images; BiFPN is then used to replace PANet for multi-scale feature fusion. In this way, small features in the images can be better extracted; Finally, the original loss function of YOLOv5 is improved, and the SIOU function is used to optimize the training model. The testing results on the NEU-DET dataset after data augmentation indicate that the improved model mAP achieves 97.2%, which is 4.1% higher than the original model, and is superior to other mainstream models. Compared to the original model, the detection speed is basically unchanged, and can quickly and accurately detect metal surface defects in real time.